intelligent edge computing and ml-based traffic classifier · 4/25/2018 · edge/fog computing...
Post on 25-May-2020
12 Views
Preview:
TRANSCRIPT
Intelligent Edge Computing and ML-based Traffic Classifier
Kwihoon Kim, Minsuk Kim (ETRI)(kwihooi@etri.re.kr, mskim16@etri.re.kr)
April 25. 2018
ITU Workshop on “Impact of AI on ICT Infrastructures“Cian, China, April 25. 2018
Agenda
• Chapter 1. Intelligent Edge Computing (IEC) for IoE services– Background– Related Work– Proposed Intelligent Edge Computing
• Chapter 2. ML-based Traffic Classifier based on Intelligent Edge Computing (IEC)– Introduction– ML-based Traffic Classifier – Conclusions
2
Chapter 1. Intelligent Edge Computing for IoE services
Background
• Internet of Things and Bigdata– Connecting things will be increased in about 20 billion by 2020
• Support connectivity to Internet
– There are lots of vertical sectors• Manufacturing, Healthcare, Energy, Smart Home, etc.
– Data will be managing, shifting through, and analyzing• Need to integrate greater computing capabilities
• Collaborated with Cloud computing– Elastic computing resources– Ubiquitous web services– Common API (RESTful)
4
Edge/Fog computing (1/2)
• Challenges– The data generated by billions of new smart sensors and devices.
• It will generate 507.5 zettabytes (1 zettabyte = 1 trillion gigabytes) of data by 2019, according to Cisco.
• Network infrastructure will be a bottleneck point between Things and Cloud
– Limitations of Cloud of Things (CoT) model• Explosive growth to various operation technologies (OT) • Smart decisions at devices• Need to time-critical application• Poor of Internet connectivity to Cloud• Data privacy, security and legal implications
5
Edge/Fog computing (2/2)
Requirements Cloud Computing Edge/Fog Computing
Latency High Low
Delay Jitter High Very low
Location of Service Within the Internet At the edge of the local network
Distance between client and server Multiple hops Minimum One hop
Security Undefined Can be defined
Attack on data enroute High probability Very low probability
Location awareness No Yes
Geo-distribution Centralized Distributed
No. of server nodes Few Very large
Support for Mobility Limited Supported
Real time interactions Supported Supported
Type of last mile connectivity Leased Line Wireless
[1] Cisco blog, http://blogs.cisco.com/perspectives/iot-from-cloud-to-fog-computing
6
Terminologies
• Cloud computing– Paradigm for enabling network access to a scalable and elastic pool of
shareable physical or virtual resources with self-service provisioning and administration on-demand. (by ITU-T Y.3500)
• Edge computing– A paradigm in which the resources of a small data center are placed at the
edge of the Internet, in close proximity to mobile devices, sensors, and end users. (by Cloudlet)
• Fog computing– A system level horizontal architecture that distributes resources and services
of computing, storage, control and networking anywhere along the continuum from Cloud to Things. (by OpenFog Consortium)
7
Data Analytic pipeline
8
Data Analytic pipeline
9
Our interested research field focus edge computing area !!
Proposed Intelligent Edge Computing
10
• A novel Domain networking architecture– Edge computing concept
• Define of specific purpose nodes
– Gateway node, networking, storing, computing, Information process
– IoE Edge Networking capability
• Constructs networking with edge entities via name-based networking (e.g., RESTful)
• Provides ingress and egress interfaces
• Supports various mobilities (consumer, producer, data)
– Intelligent Data Processing capability
• Edge analytics via AI (ML)
• Dynamic interworks with a Bigdataplatform
– Model update etc.
- Long-lived Cache server
- Information process
- Intelligent Information
Analysis
. New ML algorithm
. Clustering
. DL(CNN,RNN)
. GA, RL etc.
- External interworking
. Edge & Cloud
. Edge & other edge
- Connectivity support
. L2 connectivity
. Mobility Support
Intelligent
Edge Computing
Platform
Chapter 2. ML-based Traffic Classifier based on Intelligent Edge
Computing (IEC)
• Network Machine Learning
– Online machine learning:• A method of machine learning in which data becomes available in a
sequential order
• It is used to update our best predictor for future data at each step, as opposed to batch learning techniques which generate the best predictor by learning on the entire training data set at once
– Pipeline (software):• A chain of data-processing processes or other software entities
• The output of each element is the input of the next
12
Introduction (1/2)
[Source] https://en.wikipedia.org/wiki/Online_machine_learning
• Approach for Network Machine Learning
– Network Management:• Use multiple paths to distribute the traffic flows to avoid congestion
and ensure QoS
• ML to understand what is the relation between satisfactory QoS and the low-level metrics
• To detect the application (or attack) from network packets and to collect the history traffic data and identify a service or attack
– Data Formats:• A method of machine learning in which data becomes available in a
sequential order
• To update our best predictor for future data at each step, as opposed to batch learning techniques that generate the best prediction as following the entire data set for training
13
Introduction (2/2)
ML-based Traffic Classifier (1/7)
• Deep-learning-base Traffic Classification– Autonomously and Intelligently classifying a large amount of traffic data
– Trying to apply Deep Learning (DL) for high quality network service
14
ML-based Traffic Classifier (2/7)
15
(1) Traffic-collector (2) Data-preprocessing (3) DL-trainer/classifier (4) Monitoring
ML-based Traffic Classifier (3/7)
16
rdp – [1, 0, 0, 0, 0, 0]skype – [0, 1, 0, 0, 0, 0]python – [0, 0, 1, 0, 0, 0] slack – [0, 0, 0, 1, 0, 0]web – [0, 0, 0, 0, 1, 0]
(1) Traffic-collector (2) Data-preprocessing (3) DL-trainer/classifier (4) Monitoring
ML-based Traffic Classifier (4/7)
17
(1) Traffic-collector (2) Data-preprocessing (3) DL-trainer/classifier (4) Monitoring
ML-based Traffic Classifier (5/7)
18
(1) Traffic-collector (2) Data-preprocessing (3) DL-trainer/classifier (4) Monitoring
ML-based Traffic Classifier (6/7)
• Results for Deep-learning-based Traffic Classification– Autonomously and Intelligently classifying a large amount of traffic data
– Trying to apply Deep Learning (DL) for high quality network service
19
Data Size per Image (Packet Payload)
𝟖 × 𝟖 𝟏𝟔 × 𝟏𝟔 𝟑𝟐 × 𝟑𝟐 𝟔𝟒 × 𝟔𝟒
Accuracy(%)
Packet Size
(Bytes)
Accuracy(%)
Packet Size
(Bytes)
Accuracy(%)
Packet Size
(Bytes)
Accuracy(%)
Packet Size
(Bytes)
Data Size per Pixel
(Bit)
1 69.28 8 76.72 32 88.24 128 92.72 512
2 68.48 16 77.48 64 88.84 256 92.88 1,024
4 67.72 32 76.20 128 87.84 512 93.72 2,048
(Example)
ML-based Traffic Classifier (7/7)
• Issue of Deep-learning-based Traffic Classification– Online Learning (Streaming Data) vs Batch Learning (Batch Data)
– Need to continuously update Deep Learning Model for System Improvement
– Considering of learning data feed by streaming data and processing model evolution
20
Online Learning
Batch Learning
Conclusions
21
• Network Machine Learning
– It is required to support successful IoT and 5G services
• Difficulties of Network Machine Learning
– Lack of dataset and model
– Complexity of network behavior
– Various understanding of network machine learning
• Even though, it is the time to activate the research and standardization of network machine learning
Thank you.Questions and Comments!!
22
top related